Update src/txagent/txagent.py

src/txagent/txagent.py

@@ -12,33 +12,31 @@ from tooluniverse import ToolUniverse
 from gradio import ChatMessage
 from .toolrag import ToolRAGModel
 import torch
-# near the top of txagent.py
 import logging
+
 logger = logging.getLogger(__name__)
 logging.basicConfig(level=logging.INFO)
 
 from .utils import NoRepeatSentenceProcessor, ReasoningTraceChecker, tool_result_format
 
-
 class TxAgent:
     def __init__(self, model_name,
                  rag_model_name,
-                 tool_files_dict=None,  # None leads to the default tool files in ToolUniverse
+                 tool_files_dict=None,
                  enable_finish=True,
                  enable_rag=True,
                  enable_summary=False,
-                 init_rag_num=0,
-                 step_rag_num=10,
+                 init_rag_num=2,  # Reduced for faster initial tool selection
+                 step_rag_num=4,  # Reduced for fewer RAG calls
                  summary_mode='step',
                  summary_skip_last_k=0,
                  summary_context_length=None,
                  force_finish=True,
                  avoid_repeat=True,
                  seed=None,
-                 enable_checker=False,
+                 enable_checker=False,  # Disabled by default for speed
                  enable_chat=False,
-                 additional_default_tools=None,
-                 ):
+                 additional_default_tools=None):
         self.model_name = model_name
         self.tokenizer = None
         self.terminators = None
@@ -47,10 +45,9 @@ class TxAgent:
         self.model = None
         self.rag_model = ToolRAGModel(rag_model_name)
         self.tooluniverse = None
-        # self.tool_desc = None
-        self.prompt_multi_step = "You are a helpful assistant that will solve problems through detailed, step-by-step reasoning and actions based on your reasoning. Typically, your actions will use the provided functions. You have access to the following functions."
-        self.self_prompt = "Strictly follow the instruction."
-        self.chat_prompt = "You are helpful assistant to chat with the user."
+        self.prompt_multi_step = "You are a medical assistant solving clinical oversight issues step-by-step using provided tools."
+        self.self_prompt = "Follow instructions precisely."
+        self.chat_prompt = "You are a helpful assistant for clinical queries."
         self.enable_finish = enable_finish
         self.enable_rag = enable_rag
         self.enable_summary = enable_summary
@@ -64,7 +61,7 @@ class TxAgent:
         self.seed = seed
         self.enable_checker = enable_checker
         self.additional_default_tools = additional_default_tools
-        self.print_self_values()
+        logger.debug("TxAgent initialized with parameters: %s", self.__dict__)
 
     def init_model(self):
         self.load_models()
@@ -73,19 +70,19 @@ class TxAgent:
 
     def print_self_values(self):
         for attr, value in self.__dict__.items():
-            print(f"{attr}: {value}")
+            logger.debug("%s: %s", attr, value)
 
     def load_models(self, model_name=None):
-        if model_name is not None:
-            if model_name == self.model_name:
-                return f"The model {model_name} is already loaded."
+        if model_name is not None and model_name == self.model_name:
+            return f"The model {model_name} is already loaded."
+        if model_name:
             self.model_name = model_name
 
-        self.model = LLM(model=self.model_name)
+        self.model = LLM(model=self.model_name, dtype="float16")  # Enable FP16
         self.chat_template = Template(self.model.get_tokenizer().chat_template)
         self.tokenizer = self.model.get_tokenizer()
-
-        return f"Model {model_name} loaded successfully."
+        logger.info("Model %s loaded successfully", self.model_name)
+        return f"Model {self.model_name} loaded successfully."
 
     def load_tooluniverse(self):
         self.tooluniverse = ToolUniverse(tool_files=self.tool_files_dict)
@@ -93,9 +90,11 @@ class TxAgent:
         special_tools = self.tooluniverse.prepare_tool_prompts(
             self.tooluniverse.tool_category_dicts["special_tools"])
         self.special_tools_name = [tool['name'] for tool in special_tools]
+        logger.debug("ToolUniverse loaded with %d special tools", len(self.special_tools_name))
 
     def load_tool_desc_embedding(self):
         self.rag_model.load_tool_desc_embedding(self.tooluniverse)
+        logger.debug("Tool description embeddings loaded")
 
     def rag_infer(self, query, top_k=5):
         return self.rag_model.rag_infer(query, top_k)
@@ -109,7 +108,7 @@ class TxAgent:
             if call_agent_level >= 2:
                 call_agent = False
 
-        if not call_agent:
+        if not call_agent and self.enable_rag:
             picked_tools_prompt += self.tool_RAG(
                 message=message, rag_num=self.init_rag_num)
         return picked_tools_prompt, call_agent_level
@@ -118,293 +117,198 @@ class TxAgent:
         if conversation is None:
             conversation = []
 
-        conversation = self.set_system_prompt(
-            conversation, self.prompt_multi_step)
-        if history is not None:
-            if len(history) == 0:
-                conversation = []
-                print("clear conversation successfully")
-            else:
-                for i in range(len(history)):
-                    if history[i]['role'] == 'user':
-                        if i-1 >= 0 and history[i-1]['role'] == 'assistant':
-                            conversation.append(
-                                {"role": "assistant", "content": history[i-1]['content']})
-                        conversation.append(
-                            {"role": "user", "content": history[i]['content']})
-                    if i == len(history)-1 and history[i]['role'] == 'assistant':
-                        conversation.append(
-                            {"role": "assistant", "content": history[i]['content']})
-
+        conversation = self.set_system_prompt(conversation, self.prompt_multi_step)
+        if history:
+            conversation.extend(
+                {"role": h['role'], "content": h['content']}
+                for h in history if h['role'] in ['user', 'assistant']
+            )
         conversation.append({"role": "user", "content": message})
-
+        logger.debug("Conversation initialized with %d messages", len(conversation))
         return conversation
 
-    def tool_RAG(self, message=None,
-                 picked_tool_names=None,
-                 existing_tools_prompt=[],
-                 rag_num=5,
-                 return_call_result=False):
-        extra_factor = 30  # Factor to retrieve more than rag_num
+    def tool_RAG(self, message=None, picked_tool_names=None,
+                 existing_tools_prompt=None, rag_num=4, return_call_result=False):
+        extra_factor = 10  # Reduced from 30 for efficiency
         if picked_tool_names is None:
-            assert picked_tool_names is not None or message is not None
-            picked_tool_names = self.rag_infer(
-                message, top_k=rag_num*extra_factor)
-
-        picked_tool_names_no_special = []
-        for tool in picked_tool_names:
-            if tool not in self.special_tools_name:
-                picked_tool_names_no_special.append(tool)
-        picked_tool_names_no_special = picked_tool_names_no_special[:rag_num]
-        picked_tool_names = picked_tool_names_no_special[:rag_num]
+            picked_tool_names = self.rag_infer(message, top_k=rag_num * extra_factor)
 
+        picked_tool_names = [
+            tool for tool in picked_tool_names
+            if tool not in self.special_tools_name
+        ][:rag_num]
         picked_tools = self.tooluniverse.get_tool_by_name(picked_tool_names)
-        picked_tools_prompt = self.tooluniverse.prepare_tool_prompts(
-            picked_tools)
+        picked_tools_prompt = self.tooluniverse.prepare_tool_prompts(picked_tools)
+        logger.debug("RAG selected %d tools: %s", len(picked_tool_names), picked_tool_names)
         if return_call_result:
             return picked_tools_prompt, picked_tool_names
         return picked_tools_prompt
 
     def add_special_tools(self, tools, call_agent=False):
         if self.enable_finish:
-            tools.append(self.tooluniverse.get_one_tool_by_one_name(
-                'Finish', return_prompt=True))
-            print("Finish tool is added")
+            tools.append(self.tooluniverse.get_one_tool_by_one_name('Finish', return_prompt=True))
+            logger.debug("Finish tool added")
         if call_agent:
-            tools.append(self.tooluniverse.get_one_tool_by_one_name(
-                'CallAgent', return_prompt=True))
-            print("CallAgent tool is added")
-        else:
-            if self.enable_rag:
-                tools.append(self.tooluniverse.get_one_tool_by_one_name(
-                    'Tool_RAG', return_prompt=True))
-                print("Tool_RAG tool is added")
-
-            if self.additional_default_tools is not None:
-                for each_tool_name in self.additional_default_tools:
-                    tool_prompt = self.tooluniverse.get_one_tool_by_one_name(
-                        each_tool_name, return_prompt=True)
-                    if tool_prompt is not None:
-                        print(f"{each_tool_name} tool is added")
-                        tools.append(tool_prompt)
+            tools.append(self.tooluniverse.get_one_tool_by_one_name('CallAgent', return_prompt=True))
+            logger.debug("CallAgent tool added")
+        elif self.enable_rag:
+            tools.append(self.tooluniverse.get_one_tool_by_one_name('Tool_RAG', return_prompt=True))
+            logger.debug("Tool_RAG tool added")
+        if self.additional_default_tools:
+            for tool_name in self.additional_default_tools:
+                tool_prompt = self.tooluniverse.get_one_tool_by_one_name(tool_name, return_prompt=True)
+                if tool_prompt:
+                    tools.append(tool_prompt)
+                    logger.debug("%s tool added", tool_name)
         return tools
 
     def add_finish_tools(self, tools):
-        tools.append(self.tooluniverse.get_one_tool_by_one_name(
-            'Finish', return_prompt=True))
-        print("Finish tool is added")
+        tools.append(self.tooluniverse.get_one_tool_by_one_name('Finish', return_prompt=True))
+        logger.debug("Finish tool added")
         return tools
 
     def set_system_prompt(self, conversation, sys_prompt):
-        if len(conversation) == 0:
-            conversation.append(
-                {"role": "system", "content": sys_prompt})
+        if not conversation:
+            conversation.append({"role": "system", "content": sys_prompt})
         else:
             conversation[0] = {"role": "system", "content": sys_prompt}
         return conversation
 
-    def run_function_call(self, fcall_str,
-                          return_message=False,
-                          existing_tools_prompt=None,
-                          message_for_call_agent=None,
-                          call_agent=False,
-                          call_agent_level=None,
-                          temperature=None):
-
+    def run_function_call(self, fcall_str, return_message=False,
+                          existing_tools_prompt=None, message_for_call_agent=None,
+                          call_agent=False, call_agent_level=None, temperature=None):
         function_call_json, message = self.tooluniverse.extract_function_call_json(
             fcall_str, return_message=return_message, verbose=False)
         call_results = []
         special_tool_call = ''
-        if function_call_json is not None:
-            if isinstance(function_call_json, list):
-                for i in range(len(function_call_json)):
-                    print("\033[94mTool Call:\033[0m", function_call_json[i])
-                    if function_call_json[i]["name"] == 'Finish':
-                        special_tool_call = 'Finish'
-                        break
-                    elif function_call_json[i]["name"] == 'Tool_RAG':
-                        new_tools_prompt, call_result = self.tool_RAG(
-                            message=message,
-                            existing_tools_prompt=existing_tools_prompt,
-                            rag_num=self.step_rag_num,
-                            return_call_result=True)
-                        existing_tools_prompt += new_tools_prompt
-                    elif function_call_json[i]["name"] == 'CallAgent':
-                        if call_agent_level < 2 and call_agent:
-                            solution_plan = function_call_json[i]['arguments']['solution']
-                            full_message = (
-                                message_for_call_agent +
-                                "\nYou must follow the following plan to answer the question: " +
-                                str(solution_plan)
-                            )
-                            call_result = self.run_multistep_agent(
-                                full_message, temperature=temperature,
-                                max_new_tokens=1024, max_token=99999,
-                                call_agent=False, call_agent_level=call_agent_level)
-                            if call_result is None:
-                                call_result = "⚠️ No content returned from sub-agent."
-                            else:
-                                call_result = call_result.split('[FinalAnswer]')[-1].strip()
-                        else:
-                            call_result = "Error: The CallAgent has been disabled. Please proceed with your reasoning process to solve this question."
-                    else:
-                        call_result = self.tooluniverse.run_one_function(
-                            function_call_json[i])
-
-                    call_id = self.tooluniverse.call_id_gen()
-                    function_call_json[i]["call_id"] = call_id
-                    print("\033[94mTool Call Result:\033[0m", call_result)
-                    call_results.append({
-                        "role": "tool",
-                        "content": json.dumps({"tool_name": function_call_json[i]["name"], "content": call_result, "call_id": call_id})
-                    })
+        if function_call_json:
+            for func in function_call_json if isinstance(function_call_json, list) else [function_call_json]:
+                logger.debug("Tool Call: %s", func)
+                if func["name"] == 'Finish':
+                    special_tool_call = 'Finish'
+                    break
+                elif func["name"] == 'Tool_RAG':
+                    new_tools_prompt, call_result = self.tool_RAG(
+                        message=message, existing_tools_prompt=existing_tools_prompt,
+                        rag_num=self.step_rag_num, return_call_result=True)
+                    existing_tools_prompt += new_tools_prompt
+                elif func["name"] == 'CallAgent' and call_agent and call_agent_level < 2:
+                    solution_plan = func['arguments']['solution']
+                    full_message = (
+                        message_for_call_agent + "\nFollow this plan: " + str(solution_plan)
+                    )
+                    call_result = self.run_multistep_agent(
+                        full_message, temperature=temperature, max_new_tokens=512,
+                        max_token=2048, call_agent=False, call_agent_level=call_agent_level)
+                    call_result = call_result.split('[FinalAnswer]')[-1].strip() if call_result else "⚠️ No content from sub-agent."
+                else:
+                    call_result = self.tooluniverse.run_one_function(func)
+
+                call_id = self.tooluniverse.call_id_gen()
+                func["call_id"] = call_id
+                logger.debug("Tool Call Result: %s", call_result)
+                call_results.append({
+                    "role": "tool",
+                    "content": json.dumps({"tool_name": func["name"], "content": call_result, "call_id": call_id})
+                })
         else:
             call_results.append({
                 "role": "tool",
-                "content": json.dumps({"content": "Not a valid function call, please check the function call format."})
+                "content": json.dumps({"content": "Invalid function call format."})
             })
 
         revised_messages = [{
             "role": "assistant",
-            "content": message.strip(),
+            "content": message.strip() if message else "",
             "tool_calls": json.dumps(function_call_json)
         }] + call_results
-
-        # Yield the final result.
         return revised_messages, existing_tools_prompt, special_tool_call
 
-    def run_function_call_stream(self, fcall_str,
-                                 return_message=False,
-                                 existing_tools_prompt=None,
-                                 message_for_call_agent=None,
-                                 call_agent=False,
-                                 call_agent_level=None,
-                                 temperature=None,
-                                 return_gradio_history=True):
-
+    def run_function_call_stream(self, fcall_str, return_message=False,
+                                existing_tools_prompt=None, message_for_call_agent=None,
+                                call_agent=False, call_agent_level=None, temperature=None,
+                                return_gradio_history=True):
         function_call_json, message = self.tooluniverse.extract_function_call_json(
             fcall_str, return_message=return_message, verbose=False)
         call_results = []
         special_tool_call = ''
-        if return_gradio_history:
-            gradio_history = []
-        if function_call_json is not None:
-            if isinstance(function_call_json, list):
-                for i in range(len(function_call_json)):
-                    if function_call_json[i]["name"] == 'Finish':
-                        special_tool_call = 'Finish'
-                        break
-                    elif function_call_json[i]["name"] == 'Tool_RAG':
-                        new_tools_prompt, call_result = self.tool_RAG(
-                            message=message,
-                            existing_tools_prompt=existing_tools_prompt,
-                            rag_num=self.step_rag_num,
-                            return_call_result=True)
-                        existing_tools_prompt += new_tools_prompt
-                    elif function_call_json[i]["name"] == 'DirectResponse':
-                        call_result = function_call_json[i]['arguments']['respose']
-                        special_tool_call = 'DirectResponse'
-                    elif function_call_json[i]["name"] == 'RequireClarification':
-                        call_result = function_call_json[i]['arguments']['unclear_question']
-                        special_tool_call = 'RequireClarification'
-                    elif function_call_json[i]["name"] == 'CallAgent':
-                        if call_agent_level < 2 and call_agent:
-                            solution_plan = function_call_json[i]['arguments']['solution']
-                            full_message = (
-                                message_for_call_agent +
-                                "\nYou must follow the following plan to answer the question: " +
-                                str(solution_plan)
-                            )
-                            sub_agent_task = "Sub TxAgent plan: " + \
-                                str(solution_plan)
-                            call_result = yield from self.run_gradio_chat(
-                                full_message, history=[], temperature=temperature,
-                                max_new_tokens=1024, max_token=99999,
-                                call_agent=False, call_agent_level=call_agent_level,
-                                conversation=None,
-                                sub_agent_task=sub_agent_task)
-
-                            if call_result is not None and isinstance(call_result, str):
-                                call_result = call_result.split('[FinalAnswer]')[-1]
-                            else:
-                                call_result = "⚠️ No content returned from sub-agent."
-                        else:
-                            call_result = "Error: The CallAgent has been disabled. Please proceed with your reasoning process to solve this question."
-                    else:
-                        call_result = self.tooluniverse.run_one_function(
-                            function_call_json[i])
-
-                    call_id = self.tooluniverse.call_id_gen()
-                    function_call_json[i]["call_id"] = call_id
-                    call_results.append({
-                        "role": "tool",
-                        "content": json.dumps({"tool_name": function_call_json[i]["name"], "content": call_result, "call_id": call_id})
-                    })
-                    if return_gradio_history and function_call_json[i]["name"] != 'Finish':
-                        if function_call_json[i]["name"] == 'Tool_RAG':
-                            gradio_history.append(ChatMessage(role="assistant", content=str(call_result), metadata={
-                                                  "title": "🧰 "+function_call_json[i]['name'], "log": str(function_call_json[i]['arguments'])}))
-                        else:
-                            gradio_history.append(ChatMessage(role="assistant", content=str(call_result), metadata={
-                                                  "title": "⚒️ "+function_call_json[i]['name'], "log": str(function_call_json[i]['arguments'])}))
+        gradio_history = [] if return_gradio_history else None
+        if function_call_json:
+            for func in function_call_json if isinstance(function_call_json, list) else [function_call_json]:
+                if func["name"] == 'Finish':
+                    special_tool_call = 'Finish'
+                    break
+                elif func["name"] == 'Tool_RAG':
+                    new_tools_prompt, call_result = self.tool_RAG(
+                        message=message, existing_tools_prompt=existing_tools_prompt,
+                        rag_num=self.step_rag_num, return_call_result=True)
+                    existing_tools_prompt += new_tools_prompt
+                elif func["name"] == 'DirectResponse':
+                    call_result = func['arguments']['response']
+                    special_tool_call = 'DirectResponse'
+                elif func["name"] == 'RequireClarification':
+                    call_result = func['arguments']['unclear_question']
+                    special_tool_call = 'RequireClarification'
+                elif func["name"] == 'CallAgent' and call_agent and call_agent_level < 2:
+                    solution_plan = func['arguments']['solution']
+                    full_message = (
+                        message_for_call_agent + "\nFollow this plan: " + str(solution_plan)
+                    )
+                    sub_agent_task = "Sub TxAgent plan: " + str(solution_plan)
+                    call_result = yield from self.run_gradio_chat(
+                        full_message, history=[], temperature=temperature,
+                        max_new_tokens=512, max_token=2048, call_agent=False,
+                        call_agent_level=call_agent_level, conversation=None,
+                        sub_agent_task=sub_agent_task)
+                    call_result = call_result.split('[FinalAnswer]')[-1] if call_result else "⚠️ No content from sub-agent."
+                else:
+                    call_result = self.tooluniverse.run_one_function(func)
+
+                call_id = self.tooluniverse.call_id_gen()
+                func["call_id"] = call_id
+                call_results.append({
+                    "role": "tool",
+                    "content": json.dumps({"tool_name": func["name"], "content": call_result, "call_id": call_id})
+                })
+                if return_gradio_history and func["name"] != 'Finish':
+                    title = f"{'🧰' if func['name'] == 'Tool_RAG' else '⚒️'} {func['name']}"
+                    gradio_history.append(ChatMessage(
+                        role="assistant", content=str(call_result),
+                        metadata={"title": title, "log": str(func['arguments'])}
+                    ))
         else:
             call_results.append({
                 "role": "tool",
-                "content": json.dumps({"content": "Not a valid function call, please check the function call format."})
+                "content": json.dumps({"content": "Invalid function call format."})
             })
 
         revised_messages = [{
             "role": "assistant",
-            "content": message.strip(),
+            "content": message.strip() if message else "",
             "tool_calls": json.dumps(function_call_json)
         }] + call_results
+        return revised_messages, existing_tools_prompt, special_tool_call, gradio_history
 
-        if return_gradio_history:
-            return revised_messages, existing_tools_prompt, special_tool_call, gradio_history
-        else:
-            return revised_messages, existing_tools_prompt, special_tool_call
-
-
-    def get_answer_based_on_unfinished_reasoning(self, conversation, temperature, max_new_tokens, max_token, outputs=None):
-        if conversation[-1]['role'] == 'assisant':
+    def get_answer_based_on_unfinished_reasoning(self, conversation, temperature, max_new_tokens, max_token):
+        if conversation[-1]['role'] == 'assistant':
             conversation.append(
-                {'role': 'tool', 'content': 'Errors happen during the function call, please come up with the final answer with the current information.'})
+                {'role': 'tool', 'content': 'Errors occurred; provide final answer with current info.'})
         finish_tools_prompt = self.add_finish_tools([])
+        output = self.llm_infer(
+            messages=conversation, temperature=temperature, tools=finish_tools_prompt,
+            output_begin_string='[FinalAnswer]', max_new_tokens=max_new_tokens, max_token=max_token)
+        logger.debug("Unfinished reasoning output: %s", output)
+        return output
 
-        last_outputs_str = self.llm_infer(messages=conversation,
-                                          temperature=temperature,
-                                          tools=finish_tools_prompt,
-                                          output_begin_string='Since I cannot continue reasoning, I will provide the final answer based on the current information and general knowledge.\n\n[FinalAnswer]',
-                                          skip_special_tokens=True,
-                                          max_new_tokens=max_new_tokens, max_token=max_token)
-        print(last_outputs_str)
-        return last_outputs_str
-
-    def run_multistep_agent(self, message: str,
-                            temperature: float,
-                            max_new_tokens: int,
-                            max_token: int,
-                            max_round: int = 20,
-                            call_agent=False,
-                            call_agent_level=0) -> str:
-        """
-        Generate a streaming response using the llama3-8b model.
-        Args:
-            message (str): The input message.
-            temperature (float): The temperature for generating the response.
-            max_new_tokens (int): The maximum number of new tokens to generate.
-        Returns:
-            str: The generated response.
-        """
-        print("\033[1;32;40mstart\033[0m")
+    def run_multistep_agent(self, message: str, temperature: float, max_new_tokens: int,
+                            max_token: int, max_round: int = 10, call_agent=False, call_agent_level=0):
+        logger.debug("Starting multistep agent for message: %s", message[:100])
         picked_tools_prompt, call_agent_level = self.initialize_tools_prompt(
             call_agent, call_agent_level, message)
         conversation = self.initialize_conversation(message)
-
         outputs = []
         last_outputs = []
         next_round = True
-        function_call_messages = []
         current_round = 0
         token_overflow = False
         enable_summary = False
@@ -412,103 +316,70 @@ class TxAgent:
 
         if self.enable_checker:
             checker = ReasoningTraceChecker(message, conversation)
-        try:
-            while next_round and current_round < max_round:
-                current_round += 1
-                if len(outputs) > 0:
-                    function_call_messages, picked_tools_prompt, special_tool_call = self.run_function_call(
-                        last_outputs, return_message=True,
-                        existing_tools_prompt=picked_tools_prompt,
-                        message_for_call_agent=message,
-                        call_agent=call_agent,
-                        call_agent_level=call_agent_level,
-                        temperature=temperature)
-
-                    if special_tool_call == 'Finish':
-                        next_round = False
-                        conversation.extend(function_call_messages)
-                        if isinstance(function_call_messages[0]['content'], types.GeneratorType):
-                            function_call_messages[0]['content'] = next(
-                                function_call_messages[0]['content'])
-                        content = function_call_messages[0]['content']
-                        if content is None:
-                            return "❌ No content returned after Finish tool call."
-                        return content.split('[FinalAnswer]')[-1]
 
-                    if (self.enable_summary or token_overflow) and not call_agent:
-                        if token_overflow:
-                            print("token_overflow, using summary")
-                        enable_summary = True
-                    last_status = self.function_result_summary(
-                        conversation, status=last_status, enable_summary=enable_summary)
-
-                    if function_call_messages is not None:
-                        conversation.extend(function_call_messages)
-                        outputs.append(tool_result_format(
-                            function_call_messages))
-                    else:
-                        next_round = False
-                        conversation.extend(
-                            [{"role": "assistant", "content": ''.join(last_outputs)}])
-                        return ''.join(last_outputs).replace("</s>", "")
-                if self.enable_checker:
-                    good_status, wrong_info = checker.check_conversation()
-                    if not good_status:
-                        next_round = False
-                        print(
-                            "Internal error in reasoning: " + wrong_info)
-                        break
-                last_outputs = []
-                outputs.append("### TxAgent:\n")
-                last_outputs_str, token_overflow = self.llm_infer(messages=conversation,
-                                                                  temperature=temperature,
-                                                                  tools=picked_tools_prompt,
-                                                                  skip_special_tokens=False,
-                                                                  max_new_tokens=max_new_tokens, max_token=max_token,
-                                                                  check_token_status=True)
-                if last_outputs_str is None:
-                    print("The number of tokens exceeds the maximum limit.")
-                    if self.force_finish:
-                        return self.get_answer_based_on_unfinished_reasoning(conversation, temperature, max_new_tokens, max_token)
-                    else:
-                        return "❌ Token limit exceeded — no further steps possible."
+        while next_round and current_round < max_round:
+            current_round += 1
+            if last_outputs:
+                function_call_messages, picked_tools_prompt, special_tool_call = self.run_function_call(
+                    last_outputs, return_message=True, existing_tools_prompt=picked_tools_prompt,
+                    message_for_call_agent=message, call_agent=call_agent,
+                    call_agent_level=call_agent_level, temperature=temperature)
+
+                if special_tool_call == 'Finish':
+                    next_round = False
+                    conversation.extend(function_call_messages)
+                    content = function_call_messages[0]['content']
+                    return content.split('[FinalAnswer]')[-1] if content else "❌ No content after Finish."
+
+                if (self.enable_summary or token_overflow) and not call_agent:
+                    enable_summary = True
+                last_status = self.function_result_summary(
+                    conversation, status=last_status, enable_summary=enable_summary)
+
+                if function_call_messages:
+                    conversation.extend(function_call_messages)
+                    outputs.append(tool_result_format(function_call_messages))
                 else:
-                    last_outputs.append(last_outputs_str)
-            if max_round == current_round:
-                print("The number of rounds exceeds the maximum limit!")
-            if self.force_finish:
-                return self.get_answer_based_on_unfinished_reasoning(conversation, temperature, max_new_tokens, max_token)
-            else:
-                return None
+                    next_round = False
+                    return ''.join(last_outputs).replace("</s>", "")
+
+            if self.enable_checker:
+                good_status, wrong_info = checker.check_conversation()
+                if not good_status:
+                    logger.warning("Checker error: %s", wrong_info)
+                    break
+
+            last_outputs = []
+            last_outputs_str, token_overflow = self.llm_infer(
+                messages=conversation, temperature=temperature, tools=picked_tools_prompt,
+                max_new_tokens=max_new_tokens, max_token=max_token, check_token_status=True)
+            if last_outputs_str is None:
+                if self.force_finish:
+                    return self.get_answer_based_on_unfinished_reasoning(
+                        conversation, temperature, max_new_tokens, max_token)
+                return "❌ Token limit exceeded."
+            last_outputs.append(last_outputs_str)
 
-        except Exception as e:
-            print(f"Error: {e}")
-            if self.force_finish:
-                return self.get_answer_based_on_unfinished_reasoning(conversation, temperature, max_new_tokens, max_token)
-            else:
-                return None
+        if current_round >= max_round:
+            logger.warning("Max rounds exceeded")
+        if self.force_finish:
+            return self.get_answer_based_on_unfinished_reasoning(
+                conversation, temperature, max_new_tokens, max_token)
+        return None
 
     def build_logits_processor(self, messages, llm):
-        # Use the tokenizer from the LLM instance.
         tokenizer = llm.get_tokenizer()
         if self.avoid_repeat and len(messages) > 2:
-            assistant_messages = []
-            for i in range(1, len(messages) + 1):
-                if messages[-i]['role'] == 'assistant':
-                    assistant_messages.append(messages[-i]['content'])
-                    if len(assistant_messages) == 2:
-                        break
-            forbidden_ids = [tokenizer.encode(
-                msg, add_special_tokens=False) for msg in assistant_messages]
+            assistant_messages = [
+                m['content'] for m in messages[-3:] if m['role'] == 'assistant'
+            ][:2]
+            forbidden_ids = [tokenizer.encode(msg, add_special_tokens=False) for msg in assistant_messages]
             return [NoRepeatSentenceProcessor(forbidden_ids, 5)]
-        else:
-            return None
+        return None
 
-    def llm_infer(self, messages, temperature=0.1, tools=None,
-                  output_begin_string=None, max_new_tokens=2048,
-                  max_token=None, skip_special_tokens=True,
+    def llm_infer(self, messages, temperature=0.1, tools=None, output_begin_string=None,
+                  max_new_tokens=512, max_token=2048, skip_special_tokens=True,
                   model=None, tokenizer=None, terminators=None, seed=None, check_token_status=False):
-
         if model is None:
             model = self.model
 
@@ -516,333 +387,207 @@ class TxAgent:
         sampling_params = SamplingParams(
             temperature=temperature,
             max_tokens=max_new_tokens,
-           
             seed=seed if seed is not None else self.seed,
+            logits_processors=logits_processor
         )
 
-        prompt = self.chat_template.render(
-            messages=messages, tools=tools, add_generation_prompt=True)
-        if output_begin_string is not None:
+        prompt = self.chat_template.render(messages=messages, tools=tools, add_generation_prompt=True)
+        if output_begin_string:
             prompt += output_begin_string
 
-        if check_token_status and max_token is not None:
-            token_overflow = False
-            num_input_tokens = len(self.tokenizer.encode(
-                prompt, return_tensors="pt")[0])
-            if max_token is not None:
-                if num_input_tokens > max_token:
-                    torch.cuda.empty_cache()
-                    gc.collect()
-                    print("Number of input tokens before inference:",
-                          num_input_tokens)
-                    logger.info(
-                        "The number of tokens exceeds the maximum limit!!!!")
-                    token_overflow = True
-                    return None, token_overflow
-        output = model.generate(
-            prompt,
-            sampling_params=sampling_params,
-        )
-        output = output[0].outputs[0].text
-        print("\033[92m" + output + "\033[0m")
-        if check_token_status and max_token is not None:
-            return output, token_overflow
+        if check_token_status and max_token:
+            num_input_tokens = len(self.tokenizer.encode(prompt, return_tensors="pt")[0])
+            if num_input_tokens > max_token:
+                torch.cuda.empty_cache()
+                gc.collect()
+                logger.info("Token overflow: %d > %d", num_input_tokens, max_token)
+                return None, True
+            logger.debug("Input tokens: %d", num_input_tokens)
 
+        output = model.generate(prompt, sampling_params=sampling_params)
+        output = output[0].outputs[0].text
+        logger.debug("Inference output: %s", output[:100])
+        torch.cuda.empty_cache()  # Clear CUDA cache
+        if check_token_status:
+            return output, False
         return output
 
-    def run_self_agent(self, message: str,
-                       temperature: float,
-                       max_new_tokens: int,
-                       max_token: int) -> str:
-
-        print("\033[1;32;40mstart self agent\033[0m")
-        conversation = []
-        conversation = self.set_system_prompt(conversation, self.self_prompt)
+    def run_self_agent(self, message: str, temperature: float, max_new_tokens: int, max_token: int):
+        logger.debug("Starting self agent")
+        conversation = self.set_system_prompt([], self.self_prompt)
         conversation.append({"role": "user", "content": message})
-        return self.llm_infer(messages=conversation,
-                              temperature=temperature,
-                              tools=None,
+        return self.llm_infer(messages=conversation, temperature=temperature,
                               max_new_tokens=max_new_tokens, max_token=max_token)
 
-    def run_chat_agent(self, message: str,
-                       temperature: float,
-                       max_new_tokens: int,
-                       max_token: int) -> str:
-
-        print("\033[1;32;40mstart chat agent\033[0m")
-        conversation = []
-        conversation = self.set_system_prompt(conversation, self.chat_prompt)
+    def run_chat_agent(self, message: str, temperature: float, max_new_tokens: int, max_token: int):
+        logger.debug("Starting chat agent")
+        conversation = self.set_system_prompt([], self.chat_prompt)
         conversation.append({"role": "user", "content": message})
-        return self.llm_infer(messages=conversation,
-                              temperature=temperature,
-                              tools=None,
+        return self.llm_infer(messages=conversation, temperature=temperature,
                               max_new_tokens=max_new_tokens, max_token=max_token)
 
-    def run_format_agent(self, message: str,
-                         answer: str,
-                         temperature: float,
-                         max_new_tokens: int,
-                         max_token: int) -> str:
-
-        print("\033[1;32;40mstart format agent\033[0m")
+    def run_format_agent(self, message: str, answer: str, temperature: float, max_new_tokens: int, max_token: int):
+        logger.debug("Starting format agent")
         if '[FinalAnswer]' in answer:
             possible_final_answer = answer.split("[FinalAnswer]")[-1]
         elif "\n\n" in answer:
             possible_final_answer = answer.split("\n\n")[-1]
         else:
             possible_final_answer = answer.strip()
-        if len(possible_final_answer) == 1:
-            choice = possible_final_answer[0]
-            if choice in ['A', 'B', 'C', 'D', 'E']:
-                return choice
-        elif len(possible_final_answer) > 1:
-            if possible_final_answer[1] == ':':
-                choice = possible_final_answer[0]
-                if choice in ['A', 'B', 'C', 'D', 'E']:
-                    print("choice", choice)
-                    return choice
-
-        conversation = []
-        format_prompt = f"You are helpful assistant to transform the answer of agent to the final answer of 'A', 'B', 'C', 'D'."
-        conversation = self.set_system_prompt(conversation, format_prompt)
-        conversation.append({"role": "user", "content": message +
-                            "\nThe final answer of agent:" + answer + "\n The answer is (must be a letter):"})
-        return self.llm_infer(messages=conversation,
-                              temperature=temperature,
-                              tools=None,
+
+        if len(possible_final_answer) >= 1 and possible_final_answer[0] in ['A', 'B', 'C', 'D', 'E']:
+            return possible_final_answer[0]
+        elif len(possible_final_answer) > 1 and possible_final_answer[1] == ':' and possible_final_answer[0] in ['A', 'B', 'C', 'D', 'E']:
+            return possible_final_answer[0]
+
+        conversation = self.set_system_prompt(
+            [], "Transform the answer to a single letter: 'A', 'B', 'C', 'D', or 'E'.")
+        conversation.append({"role": "user", "content": f"Original: {message}\nAnswer: {answer}\nFinal answer (letter):"})
+        return self.llm_infer(messages=conversation, temperature=temperature,
                               max_new_tokens=max_new_tokens, max_token=max_token)
 
-    def run_summary_agent(self, thought_calls: str,
-                          function_response: str,
-                          temperature: float,
-                          max_new_tokens: int,
-                          max_token: int) -> str:
-        print("\033[1;32;40mSummarized Tool Result:\033[0m")
-        generate_tool_result_summary_training_prompt = """Thought and function calls: 
-{thought_calls}
-Function calls' responses:
-\"\"\"
-{function_response}
-\"\"\"
-Based on the Thought and function calls, and the function calls' responses, you need to generate a summary of the function calls' responses that fulfills the requirements of the thought. The summary MUST BE ONE sentence and include all necessary information.
-Directly respond with the summarized sentence of the function calls' responses only. 
-Generate **one summarized sentence** about "function calls' responses" with necessary information, and respond with a string:
-            """.format(thought_calls=thought_calls, function_response=function_response)
-        conversation = []
-        conversation.append(
-            {"role": "user", "content": generate_tool_result_summary_training_prompt})
-        output = self.llm_infer(messages=conversation,
-                                temperature=temperature,
-                                tools=None,
+    def run_summary_agent(self, thought_calls: str, function_response: str,
+                         temperature: float, max_new_tokens: int, max_token: int):
+        logger.debug("Starting summary agent")
+        prompt = f"""Thought and function calls: {thought_calls}
+Function responses: {function_response}
+Summarize the function responses in one sentence with all necessary information."""
+        conversation = [{"role": "user", "content": prompt}]
+        output = self.llm_infer(messages=conversation, temperature=temperature,
                                 max_new_tokens=max_new_tokens, max_token=max_token)
-
         if '[' in output:
             output = output.split('[')[0]
         return output
 
     def function_result_summary(self, input_list, status, enable_summary):
-        """
-        Processes the input list, extracting information from sequences of 'user', 'tool', 'assistant' roles.
-        Supports 'length' and 'step' modes, and skips the last 'k' groups.
-        Parameters:
-            input_list (list): A list of dictionaries containing role and other information.
-            summary_skip_last_k (int): Number of groups to skip from the end. Defaults to 0.
-            summary_context_length (int): The context length threshold for the 'length' mode.
-            last_processed_index (tuple or int): The last processed index.
-        Returns:
-            list: A list of extracted information from valid sequences.
-        """
         if 'tool_call_step' not in status:
             status['tool_call_step'] = 0
+        if 'step' not in status:
+            status['step'] = 0
+        status['step'] += 1
 
         for idx in range(len(input_list)):
-            pos_id = len(input_list)-idx-1
-            if input_list[pos_id]['role'] == 'assistant':
-                if 'tool_calls' in input_list[pos_id]:
-                    if 'Tool_RAG' in str(input_list[pos_id]['tool_calls']):
-                        status['tool_call_step'] += 1
+            pos_id = len(input_list) - idx - 1
+            if input_list[pos_id]['role'] == 'assistant' and 'tool_calls' in input_list[pos_id]:
+                if 'Tool_RAG' in str(input_list[pos_id]['tool_calls']):
+                    status['tool_call_step'] += 1
                 break
 
-        if 'step' in status:
-            status['step'] += 1
-        else:
-            status['step'] = 0
-
         if not enable_summary:
             return status
 
         if 'summarized_index' not in status:
             status['summarized_index'] = 0
-
         if 'summarized_step' not in status:
             status['summarized_step'] = 0
-
         if 'previous_length' not in status:
             status['previous_length'] = 0
-
         if 'history' not in status:
             status['history'] = []
 
-        function_response = ''
-        idx = 0
-        current_summarized_index = status['summarized_index']
-
-        status['history'].append(self.summary_mode == 'step' and status['summarized_step']
-                                 < status['step']-status['tool_call_step']-self.summary_skip_last_k)
+        status['history'].append(
+            self.summary_mode == 'step' and status['summarized_step'] < status['step'] - status['tool_call_step'] - self.summary_skip_last_k)
 
-        idx = current_summarized_index
+        idx = status['summarized_index']
+        function_response = ''
+        this_thought_calls = None
         while idx < len(input_list):
-            if (self.summary_mode == 'step' and status['summarized_step'] < status['step']-status['tool_call_step']-self.summary_skip_last_k) or (self.summary_mode == 'length' and status['previous_length'] > self.summary_context_length):
-
+            if (self.summary_mode == 'step' and status['summarized_step'] < status['step'] - status['tool_call_step'] - self.summary_skip_last_k) or \
+               (self.summary_mode == 'length' and status['previous_length'] > self.summary_context_length):
                 if input_list[idx]['role'] == 'assistant':
                     if 'Tool_RAG' in str(input_list[idx]['tool_calls']):
                         this_thought_calls = None
                     else:
-                        if len(function_response) != 0:
-                            print("internal summary")
+                        if function_response:
                             status['summarized_step'] += 1
                             result_summary = self.run_summary_agent(
-                                thought_calls=this_thought_calls,
-                                function_response=function_response,
-                                temperature=0.1,
-                                max_new_tokens=1024,
-                                max_token=99999
-                            )
-
+                                thought_calls=this_thought_calls, function_response=function_response,
+                                temperature=0.1, max_new_tokens=512, max_token=2048)
                             input_list.insert(
-                                last_call_idx+1, {'role': 'tool', 'content': result_summary})
+                                last_call_idx + 1, {'role': 'tool', 'content': result_summary})
                             status['summarized_index'] = last_call_idx + 2
                             idx += 1
-
                         last_call_idx = idx
-                        this_thought_calls = input_list[idx]['content'] + \
-                            input_list[idx]['tool_calls']
+                        this_thought_calls = input_list[idx]['content'] + input_list[idx]['tool_calls']
                         function_response = ''
-
-                elif input_list[idx]['role'] == 'tool' and this_thought_calls is not None:
+                elif input_list[idx]['role'] == 'tool' and this_thought_calls:
                     function_response += input_list[idx]['content']
                     del input_list[idx]
                     idx -= 1
-
             else:
                 break
             idx += 1
 
-        if len(function_response) != 0:
+        if function_response:
             status['summarized_step'] += 1
             result_summary = self.run_summary_agent(
-                thought_calls=this_thought_calls,
-                function_response=function_response,
-                temperature=0.1,
-                max_new_tokens=1024,
-                max_token=99999
-            )
-
+                thought_calls=this_thought_calls, function_response=function_response,
+                temperature=0.1, max_new_tokens=512, max_token=2048)
             tool_calls = json.loads(input_list[last_call_idx]['tool_calls'])
             for tool_call in tool_calls:
                 del tool_call['call_id']
             input_list[last_call_idx]['tool_calls'] = json.dumps(tool_calls)
             input_list.insert(
-                last_call_idx+1, {'role': 'tool', 'content': result_summary})
+                last_call_idx + 1, {'role': 'tool', 'content': result_summary})
             status['summarized_index'] = last_call_idx + 2
 
         return status
 
-    # Following are Gradio related functions
-
-    # General update method that accepts any new arguments through kwargs
     def update_parameters(self, **kwargs):
+        updated_attributes = {}
         for key, value in kwargs.items():
             if hasattr(self, key):
                 setattr(self, key, value)
-
-        # Return the updated attributes
-        updated_attributes = {key: value for key,
-                              value in kwargs.items() if hasattr(self, key)}
+                updated_attributes[key] = value
+        logger.debug("Updated parameters: %s", updated_attributes)
         return updated_attributes
 
-    def run_gradio_chat(self, message: str,
-                        history: list,
-                        temperature: float,
-                        max_new_tokens: int,
-                        max_token: int,
-                        call_agent: bool,
-                        conversation: gr.State,
-                        max_round: int = 20,
-                        seed: int = None,
-                        call_agent_level: int = 0,
-                        sub_agent_task: str = None,
-                        uploaded_files: list = None) -> str:
-        """
-        Generate a streaming response using the loaded model.
-        Args:
-            message (str): The input message (with file content if uploaded).
-            history (list): The conversation history used by ChatInterface.
-            temperature (float): Sampling temperature.
-            max_new_tokens (int): Max new tokens.
-            max_token (int): Max total tokens allowed.
-        Returns:
-            str: Final assistant message.
-        """
-        logger.debug(f"[TxAgent] Chat started, message: {message[:100]}...")
-        print("\033[1;32;40m[TxAgent] Chat started\033[0m")
-
+    def run_gradio_chat(self, message: str, history: list, temperature: float,
+                        max_new_tokens: int, max_token: int, call_agent: bool,
+                        conversation: gr.State, max_round: int = 10, seed: int = None,
+                        call_agent_level: int = 0, sub_agent_task: str = None,
+                        uploaded_files: list = None):
+        logger.debug("Chat started, message: %s", message[:100])
         if not message or len(message.strip()) < 5:
             yield "Please provide a valid message or upload files to analyze."
-            return "Invalid input."
+            return
 
         if message.startswith("[\U0001f9f0 Tool_RAG") or message.startswith("⚒️"):
-            return ""
-
-        outputs = []
-        outputs_str = ''
-        last_outputs = []
+            return
 
         picked_tools_prompt, call_agent_level = self.initialize_tools_prompt(
-            call_agent,
-            call_agent_level,
-            message)
-
+            call_agent, call_agent_level, message)
         conversation = self.initialize_conversation(
-            message,
-            conversation=conversation,
-            history=history)
+            message, conversation, history)
         history = []
 
         next_round = True
-        function_call_messages = []
         current_round = 0
         enable_summary = False
         last_status = {}
         token_overflow = False
 
         if self.enable_checker:
-            checker = ReasoningTraceChecker(
-                message, conversation, init_index=len(conversation))
+            checker = ReasoningTraceChecker(message, conversation, init_index=len(conversation))
 
         try:
             while next_round and current_round < max_round:
                 current_round += 1
-                logger.debug(f"Round {current_round}, conversation length: {len(conversation)}")
-
+                last_outputs = []
                 if last_outputs:
                     function_call_messages, picked_tools_prompt, special_tool_call, current_gradio_history = yield from self.run_function_call_stream(
-                        last_outputs, return_message=True,
-                        existing_tools_prompt=picked_tools_prompt,
-                        message_for_call_agent=message,
-                        call_agent=call_agent,
-                        call_agent_level=call_agent_level,
-                        temperature=temperature)
-
+                        last_outputs, return_message=True, existing_tools_prompt=picked_tools_prompt,
+                        message_for_call_agent=message, call_agent=call_agent,
+                        call_agent_level=call_agent_level, temperature=temperature)
                     history.extend(current_gradio_history)
 
-                    if special_tool_call == 'Finish' and function_call_messages:
+                    if special_tool_call == 'Finish':
                         yield history
                         next_round = False
                         conversation.extend(function_call_messages)
                         return function_call_messages[0]['content']
 
-                    elif special_tool_call in ['RequireClarification', 'DirectResponse']:
+                    if special_tool_call in ['RequireClarification', 'DirectResponse']:
                         last_msg = history[-1] if history else ChatMessage(role="assistant", content="Response needed.")
                         history.append(ChatMessage(role="assistant", content=last_msg.content))
                         yield history
@@ -851,64 +596,46 @@ Generate **one summarized sentence** about "function calls' responses" with nece
 
                     if (self.enable_summary or token_overflow) and not call_agent:
                         enable_summary = True
-
                     last_status = self.function_result_summary(
-                        conversation, status=last_status,
-                        enable_summary=enable_summary)
+                        conversation, status=last_status, enable_summary=enable_summary)
 
                     if function_call_messages:
                         conversation.extend(function_call_messages)
                         yield history
                     else:
                         next_round = False
-                        conversation.append({"role": "assistant", "content": ''.join(last_outputs)})
                         return ''.join(last_outputs).replace("</s>", "")
 
                 if self.enable_checker:
                     good_status, wrong_info = checker.check_conversation()
                     if not good_status:
-                        print("Checker flagged reasoning error: ", wrong_info)
+                        logger.warning("Checker error: %s", wrong_info)
                         break
 
-                last_outputs = []
                 last_outputs_str, token_overflow = self.llm_infer(
-                    messages=conversation,
-                    temperature=temperature,
-                    tools=picked_tools_prompt,
-                    skip_special_tokens=False,
-                    max_new_tokens=max_new_tokens,
-                    max_token=max_token,
-                    seed=seed,
-                    check_token_status=True)
-
-                logger.debug(f"llm_infer output: {last_outputs_str[:100] if last_outputs_str else None}, token_overflow: {token_overflow}")
+                    messages=conversation, temperature=temperature, tools=picked_tools_prompt,
+                    max_new_tokens=max_new_tokens, max_token=max_token, seed=seed, check_token_status=True)
 
                 if last_outputs_str is None:
-                    logger.warning("llm_infer returned None due to token overflow")
                     if self.force_finish:
                         last_outputs_str = self.get_answer_based_on_unfinished_reasoning(
                             conversation, temperature, max_new_tokens, max_token)
                         history.append(ChatMessage(role="assistant", content=last_outputs_str.strip()))
                         yield history
                         return last_outputs_str
-                    else:
-                        error_msg = "Token limit exceeded. Please reduce input size or increase max_token."
-                        history.append(ChatMessage(role="assistant", content=error_msg))
-                        yield history
-                        return error_msg
+                    error_msg = "Token limit exceeded."
+                    history.append(ChatMessage(role="assistant", content=error_msg))
+                    yield history
+                    return error_msg
 
                 last_thought = last_outputs_str.split("[TOOL_CALLS]")[0]
-
                 for msg in history:
-                    if msg.metadata is not None:
+                    if msg.metadata:
                         msg.metadata['status'] = 'done'
 
                 if '[FinalAnswer]' in last_thought:
                     parts = last_thought.split('[FinalAnswer]', 1)
-                    if len(parts) == 2:
-                        final_thought, final_answer = parts
-                    else:
-                        final_thought, final_answer = last_thought, ""
+                    final_thought, final_answer = parts if len(parts) == 2 else (last_thought, "")
                     history.append(ChatMessage(role="assistant", content=final_thought.strip()))
                     yield history
                     history.append(ChatMessage(role="assistant", content="**🧠 Final Analysis:**\n" + final_answer.strip()))
@@ -919,45 +646,28 @@ Generate **one summarized sentence** about "function calls' responses" with nece
 
                 last_outputs.append(last_outputs_str)
 
-            if next_round:
-                if self.force_finish:
-                    last_outputs_str = self.get_answer_based_on_unfinished_reasoning(
-                        conversation, temperature, max_new_tokens, max_token)
-                    if '[FinalAnswer]' in last_outputs_str:
-                        parts = last_outputs_str.split('[FinalAnswer]', 1)
-                        if len(parts) == 2:
-                            final_thought, final_answer = parts
-                        else:
-                            final_thought, final_answer = last_outputs_str, ""
-                        history.append(ChatMessage(role="assistant", content=final_thought.strip()))
-                        yield history
-                        history.append(ChatMessage(role="assistant", content="**🧠 Final Analysis:**\n" + final_answer.strip()))
-                        yield history
-                    else:
-                        history.append(ChatMessage(role="assistant", content=last_outputs_str.strip()))
-                        yield history
-                else:
-                    yield "The number of reasoning rounds exceeded the limit."
+            if next_round and self.force_finish:
+                last_outputs_str = self.get_answer_based_on_unfinished_reasoning(
+                    conversation, temperature, max_new_tokens, max_token)
+                parts = last_outputs_str.split('[FinalAnswer]', 1)
+                final_thought, final_answer = parts if len(parts) == 2 else (last_outputs_str, "")
+                history.append(ChatMessage(role="assistant", content=final_thought.strip()))
+                yield history
+                history.append(ChatMessage(role="assistant", content="**🧠 Final Analysis:**\n" + final_answer.strip()))
+                yield history
 
         except Exception as e:
-            logger.error(f"Exception in run_gradio_chat: {e}", exc_info=True)
-            error_msg = f"An error occurred: {e}"
+            logger.error("Exception in run_gradio_chat: %s", e, exc_info=True)
+            error_msg = f"Error: {e}"
             history.append(ChatMessage(role="assistant", content=error_msg))
             yield history
             if self.force_finish:
                 last_outputs_str = self.get_answer_based_on_unfinished_reasoning(
                     conversation, temperature, max_new_tokens, max_token)
-                if '[FinalAnswer]' in last_outputs_str:
-                    parts = last_outputs_str.split('[FinalAnswer]', 1)
-                    if len(parts) == 2:
-                        final_thought, final_answer = parts
-                    else:
-                        final_thought, final_answer = last_outputs_str, ""
-                    history.append(ChatMessage(role="assistant", content=final_thought.strip()))
-                    yield history
-                    history.append(ChatMessage(role="assistant", content="**🧠 Final Analysis:**\n" + final_answer.strip()))
-                    yield history
-                else:
-                    history.append(ChatMessage(role="assistant", content=last_outputs_str.strip()))
-                    yield history
+                parts = last_outputs_str.split('[FinalAnswer]', 1)
+                final_thought, final_answer = parts if len(parts) == 2 else (last_outputs_str, "")
+                history.append(ChatMessage(role="assistant", content=final_thought.strip()))
+                yield history
+                history.append(ChatMessage(role="assistant", content="**🧠 Final Analysis:**\n" + final_answer.strip()))
+                yield history
             return error_msg